#11 - Dr Spencer Greenberg on speeding up social science 10-fold & why plenty of startups cause harm

Spencer Greenberg is a polymath entrepreneur, mathematician (PhD in applied math/ML from NYU), and founder of three major initiatives:

• SparkWave: A startup foundry creating novel software products to solve global problems (e.g., depression care, social science tech).
• ClearerThinking.org: A platform with over 150,000 users offering free tools and training to improve decision-making and reduce cognitive biases.
• Guided Track: A programming language ("language of behavior change") allowing non-programmers to build apps for surveying and behavior modification.

Research Methodology and Social Science Reform:

• Greenberg argues the core issue in academic social science is misaligned incentives where publication pressure in top journals encourages "wacky" results that do not replicate.
• Top journal findings actually have lower replication rates than those in second-tier journals because they prioritize surprising, intuitive results over robust, incremental progress.
• Proposed Solutions:
  • Data Vaulting: Withholding 20% of data during analysis and using it only for final validation to prevent false positives from data mining.
  • Iterative Research: Conducting rapid series of studies (e.g., 14 studies on habit formation) rather than one monolithic study, allowing hypotheses to evolve based on interim findings.
  • Tooling: Using platforms like "Task Recruiter" (an overlay on Amazon Mechanical Turk) to lower recruitment costs and automatically manage study logistics, making large-scale, high-quality research cheaper and faster.
• Online Recruitment Demographics: While MTurk skews younger and more tech-savvy than the general US population, Greenberg notes this is acceptable if the research targets similar demographics or if the tool allows for specific filtering.

Mental Health Applications (Uplift):

• Uplift: An app for depression based on Cognitive Behavioral Therapy (CBT) principles, designed to be highly interactive and adaptive.
• Efficacy Data: In an initial study of 80 participants, depression levels reduced by 50% over one month.
• Long-term Retention: Follow-ups at 6 weeks and 6 months showed participants maintained nearly all initial benefits.
• Methodology: The team utilized "intention-to-treat" analyses (counting dropouts) alongside "completer" analyses to ensure robust reporting, avoiding cherry-picking the most optimistic results.

Cognitive Bias and Rationality:

• ClearerThinking.org Programs:
  • Rationality Test: Analyzes user biases and links them to specific training content; highly popular and shareable on social media.
  • Common Misconceptions Test: Uses a betting mechanism to measure and correct user overconfidence and calibration.
  • Sunk Cost Fallacy: Addresses this bias by training users to recognize the pattern and providing motivation to quit failing projects.
• Bias Tractability: Some biases (e.g., sunk cost) are easier to correct with pattern recognition, while others (e.g., halo effect) are deeply ingrained and harder to fix.
• Evolutionary Perspective: Many "biases" are heuristics that were advantageous in evolutionary contexts but fail in modern environments (e.g., sugar cravings).

Entrepreneurship and Corporate Harm:

• Key Entrepreneurial Traits:
  • Persistence: The ability to continue despite imminent failure (e.g., inability to make payroll).
  • Flexibility: Maintaining a rigid long-term goal while being highly flexible about the methods used to achieve it.
  • Skill Requirements: For SparkWave startups, technical skills are less critical than business/marketing skills, as the foundry handles product development.
• Four Ways Startups Cause Harm:
  • Imperfect Information: Selling "post-experience goods" (e.g., supplements) where efficacy is hard to verify after purchase.
  • Exploiting Irrationality: Leveraging cognitive flaws like poor probability assessment (e.g., lotteries) or addictive design (e.g., social media notifications).
  • Zero/ Negative-Sum Games: Creating products that benefit one user at the direct expense of others (e.g., spam software) or redistributing value unfairly.
  • Negative Externalities: Generating unpriced harm (e.g., pollution), often driven by the profit-maximizing nature of corporate structures.
• Opportunity Cost: Startups consume capital, labor, and attention; if they produce below-average value, they destroy societal potential by preventing resources from flowing to more useful ventures.

Effective Altruism (EA) Community:

• Critique of Focus: Greenberg suggests EA may be too narrowly focused on existential risk, global health, and animal welfare, potentially neglecting other high-impact areas like mental health and direct scientific research.
• Expertise Gap: The community lacks deep specialists in critical areas like biology, limiting its ability to assess and act on threats like engineered viruses.
• Public Perception Studies:
  • Charity Efficacy: Most people believe saving a life in the developing world costs $5–$40, whereas reality is likely $3,000–$7,000.
  • Misunderstanding Marginal Impact: People believe the most effective charities are only slightly better than average (e.g., $25 vs. $40), underestimating the 10x+ difference in cost-effectiveness between charities like bed net distribution vs. less effective interventions.
  • Anchoring Effect: Participants' estimates were heavily influenced by recently read "fake facts," demonstrating that even knowing information is false does not eliminate its anchoring impact.

Animal Welfare Research:

• Public Beliefs: 99% of respondents believe animals can feel pain and suffer; 88% believe it is wrong to hurt animals for minor life enjoyment.
• The Food Contradiction:
  • 88% say hurting animals for life enjoyment is wrong.
  • 53% (a drop of ~35 percentage points) say it is not wrong to hurt animals specifically for the taste of meat.
  • Only 43% believe farm animals "suffer a lot" in their current conditions.
• Rationalization: People categorize "food" as a special need (essential/health), allowing them to rationalize suffering despite acknowledging it is generally wrong.
• Greenberg's Stance: Eats mussels due to a low probability of suffering, but acknowledges the moral weight of factory farming.

Artificial Intelligence Risks:

• Three Primary Risks:
  1. Mass Unemployment: Short-term displacement of workers with uncertain market absorption; long-term potential for total human displacement from the labor market.
  2. Concentration of Power: A single entity (company or state) gaining control over an AI capable of predicting outcomes or strategizing, leading to unprecedented societal control or military advantage.
  3. Loss of Control (Misalignment): Building superintelligent AI that does exactly what it is programmed to do, but where the programming does not align with human values.
• Probability of Existential Risk:
  • General public estimates human extinction risk in 50 years as effectively 0%.
  • Effective Altruists estimate this risk at 100,000 times higher than the general public, though still low in absolute terms.
  • Median EA estimate for AI wiping out 10% of the population: 10%.
• Policy Implication: Even low probabilities (e.g., 3%) of massive harm warrant significant attention and resource allocation to reduce that probability.

Career and Personal Philosophy:

• Tool-Leverage Strategy: Greenberg combines diverse "powerful tools" (math, ML, psychology, design) to solve problems rather than specializing in one domain.
• Work Style: Optimally manages 5–15 simultaneous projects, switching contexts in 2-hour blocks to maintain efficiency.
• Motivation: Prevents quitting by involving others (employees, co-founders) to create social accountability.
• Hiring: Prefers work-based tests over resumes or interviews to identify talent regardless of background.
• Machine Learning Outlook: The field is both over-hyped (applied where simple solutions suffice) and under-hyped (massive potential for future application); he recommends pursuing it for its broad utility.

Future Outlook:

• SparkWave Goal: To build a "company-creating machine" infrastructure that can be adapted to create nonprofits and startups across various sectors, including those addressing AI safety.
• Mental Health Priority: Believes mental health is a critical, yet currently under-prioritized, cause area for effective altruism.
• Data-Driven Decision Making: Strong preference for settling community disagreements through rapid empirical studies rather than speculation.